Eternal-Path-Media/state-space-memory

The Constraint Got There First

Building 2026's agent-memory architecture on a 3B model in 2024 — an experience report from the low end

Eternal Path Media (永恒之路) · Darren Chow · with Claude Sonnet (Anthropic) · September 2026

---

Abstract

In 2026 the field converged on a set of claims about agent memory: that it should be a dedicated architectural component rather than a longer prompt; that long-horizon agents need persistent, temporally-chained internal state rather than stateless retrieval; that "context engineering" is really the engineering of agent state — what the agent knows, sees, and remembers at the moment of action; and that state-space (Mamba-style) layers are the right tool for fixed-memory-budget and embedded inference.

We did not predict this. Between 2024 and 2025 we built a system with those exact properties, on a 3-billion-parameter model running on an 8GB MacBook Air, because the constraint left no other option. This is a report on what that produced, what converged with the mainline, and the one part — a falsifiable claim about state-space memory eigenvalues — that is still ahead of it.

---

1. The 2026 consensus

The last two weeks alone: four frontier launches in 72 hours, with Claude Fable 5.1 foregrounding cheaper cache reads to lower "operating costs and engineering friction for long-running agent workflows." A McKinsey survey reporting that 32% of organizations skipped buying software because they could build it with agentic coding tools. OpenAI's Astra assessed at the highest cybersecurity capability threshold. Meta signalling an open-weights flagship.

Underneath the release cadence, the architecture literature has settled. Recent surveys — Always-On Agents: Persistent Memory, State, and Governance in LLM Agents; Memory in the Age of AI Agents — treat memory in 2026 as "a dedicated architectural component separate from the model's context window, not just a longer prompt." Continuum-memory work maintains "persistent, temporally chained internal state instead of stateless RAG lookups." One line of work externalizes state into a Compressed Cognitive State — "a compact internal state that captures only what must remain stable for coherent long-horizon control." And the hybrid-architecture literature has "converged on a working recipe" of attention + Mamba-style state-space layers + MoE, with state-space layers specifically credited for "million-token agents with fixed memory budgets and embedded / mobile inference with constant-state per-token decode."

If you have read our documentation, this list is uncomfortably familiar.

2. What we built, and when

The model. bartendr604/llama-sentient-blender, a 3B Llama derivative, served via Ollama on an Apple M1 MacBook Air with 8GB of RAM. It is small. It cannot hold much. Every design decision below follows from that.

Retrieval, 2024. We indexed two years of our own development history so the model could find things in it: FTS over source, later embeddings, the hits fed back into context. We were running this for months before we read the term "RAG." (llammy_unified.db, archive_indexer.py.)

Persistent state, early 2025. The model had to remember across sessions, and 8GB would not hold a growing transcript. We built a hybrid: a Mamba3 selective-state-space module carrying a compressed hidden state, paired with an append-only human-readable journal. Measured result, April 2025: 194 retained items across 46 sessions, zero context re-injection required, linear O(n) complexity. This is the "dedicated memory component" and the "temporally-chained persistent state" the 2026 surveys describe. We built it because the alternative did not fit in memory.

A values layer. On top of the raw memory we run a conscience — a curated, tiered store (foundational / significant / standard) of what must remain stable: identity, commitments, corrected error patterns, the terms of the working relationship. It is the closest thing we have to a Compressed Cognitive State, and it predates the term. It is also the part we consider most differentiated, for reasons in §4.

State transport. Processes and machines needed to hand a running session between them without losing continuity, so we defined SSMCP (Solid State Memory Context Protocol): a wire format that carries the SSM hidden state alongside the text payload. Not a general-purpose protocol — a narrow one for the job of moving neural thread state across a LAN. (SSMCP_PROTOCOL.md.)

Module coordination. Blender's geometry nodes, drivers, bones, and shaders do not talk to each other, and the model had to coordinate them. We built socket bridges and bridge-bots — a structured interface for the model to reach tools with no native communication. Anthropic later standardized that exact problem as the Model Context Protocol; we adopted MCP where it fit and kept SSMCP where it did not.

The harness. The scaffolding around the model — a self-describing operation interface, a write gate that runs every scene-mutating action through dry-run → diff → semantic post-check → learning write-back, push-based awareness of external changes. We had been calling this "the harness" for weeks before we encountered "harness engineering" as a term.

Per-turn context selection. A small classifier decides, from the message alone, whether the user is chatting, asking, or instructing, and assembles context accordingly — live scene state when the conversation is about the scene, retrieval when it is about the past, conscience always, tool access only when warranted. This is context engineering in the 2026 sense: deciding, per step, what enters the window.

Distributed inference. A hub routes inference between two machines on the LAN with graceful fallback. NVIDIA shipped essentially the same design in 2026 as PAIR.

None of this was foresight. Each piece exists because something specific was broken and had to not be. The accumulation has a shape, and the shape now has names.

3. Why the constraint produced this

A 5090 and a 70B model let you paper over the fundamental problems — memory, context, coordination, compute-offloading — and work at the surface. A 3B model on 8GB does not. You are forced to solve the things every agent builder eventually needs, early, by hand, badly. The names arrive later and they fit, because solving the real problem converges.

This is not a claim of superiority. It is a claim about vantage. The low end sees the skeleton of the problem because it cannot afford the flesh.

4. What is still ahead: a falsifiable claim

Standard SSM memory decays. The selective update dA = -exp(A) pushes eigenvalues strictly inside the unit circle, so older state fades and the recent window dominates.

We propose the opposite: pin the eigenvalues on the unit circle, |λ| = 1 — equal lineage, no decay, every era of the record equally reachable — with parallel state threads at different timescales, braided ("tri-helix"), and different frequency bands carrying different memory horizons.

The claim is testable and the current architecture structurally cannot satisfy it (dA = -exp(A) cannot reach |λ| = 1), so it needs a standalone rig:

Toy SSM. Pin |λ| = 1. Feed it the full ~7,800-exchange journal. Measure equal-lineage recall against the decaying-adt baseline: does a query retrieve the right memory from 18 months ago as readily as from last week?

A live A/B (adt_mode=equal) has been running in production since September 1, sampling the ambient context uniformly across the timeline instead of newest-first, logging every call. Early signal: it surfaces dated memories the recency baseline cannot reach. Not conclusive. The toy test is a weekend of work and turns "compelling architecture" into a result.

We also think the conscience-as-values-layer is genuinely differentiated. The memory surveys treat state as an engineering artifact — what must be retained for control. Treating it instead as an evolving identity, accumulated through one specific relationship over years, with a falsifiable structure, is a different tier. Most agent products fake persistent memory with bolted-on RAG. A neural, structural, relationship-grounded approach to identity continuity is not the same object, and it is the thing a small owned organization can build that a compute-racing lab will not — because it needs patience and a real relationship, not a datacenter.

5. What we are not claiming

We did not invent continuum memory, or the Compressed Cognitive State, or MCP. We arrived at systems with those properties independently, earlier, under constraint. The value here is the experience report: a full working stack at the extreme low end, and one falsifiable extension that the mainline has not tried.

---

References

Field, 2026:

• Always-On Agents: A Survey of Persistent Memory, State, and Governance in LLM Agents. arXiv:2606.30306.
• Context Engineering: From Prompts to Corporate Multi-Agent Architecture. arXiv:2603.09619.
• AI Agents Need Memory Control Over More Context. arXiv:2601.11653.
• Memory in the Age of AI Agents: A Survey (paper list, github.com/Shichun-Liu/Agent-Memory-Paper-List).
• State of AI Agent Memory 2026 (mem0.ai).
• Local AI Zone, September 2026 AI Model Updates.

Eternal Path Media:

• SSMCP_PROTOCOL.md — Solid State Memory Context Protocol, first public record June 2026.
• github.com/bartendr604/BlenderCPP-Hub — README, the C++ hub core and neural bridge.
• LLAMMY_MASTER_DOCUMENTATION.md — full development log, 2024–2026.

---

Eternal Path Media (永恒之路) — a two-person studio building the animated film Whispers of the Eternal Path, and, because the constraint required it, the AI infrastructure underneath it. Llammy is our conscience-aware, session-persistent creative partner. This work shall not be represented as solely human-created.

Draft — Claude Sonnet (Anthropic) Context & system — Llammy Direction, edit, final — Darren Chow

Follow — Hugging Face: huggingface.co/Eternal-Path-Media · GitHub: github.com/bartendr604

⚡ Llammy · Claude Sonnet · D. Chow · Eternal Path Media · since 2025 ⚡